1. NY, 14 December 2007 Emmanuel Lesaffre Biostatistical Centre, K.U.Leuven, Leuven, Belgium Dept of Biostatistics, Erasmus MC, Rotterdam, the Netherlands Clinical Research Methodology Course Randomized Clinical Trials and the “REAL WORLD” Superiority, equivalence and non-inferiority trials

3. 1. Review of study designs Type of null-hypothesis: Superiority Equivalence Non-inferiority Here, only two-group comparisons: E(experimental) & C(ontrol)

4. 1. Review of study designs Superiority trial: most classical design Prove that E is better than C Equivalence trial: showing bio-equivalence Show that E is equivalent to C Non-inferiority trial: popular in active –controlled trials Show that E is not (much) worse than C

5. Example I: GUSTO-I study Comparison of two thrombolytic drugs: SK (C) and rt-PA (E) Primary endpoint = 30-day mortality (binary) SK: 10,370 patientsrt-PA: 10,348 patients SK: 7.4% rt-PA: 6.3% H 0 :  = 0 Chi-square test: 8.94P=0.0028 95% C.I. for  : [0.36%, 1.73%] 2. Superiority trial

6. Example II: part of GUSTO-I Comparison of two thrombolytic drugs: SK and rt-PA Primary endpoint = 30-day mortality (binary) SK: 1,000 patientsrt-PA: 1,000 patients SK: 7.4% rt-PA: 6.3% H 0 :  = 0 Chi-square test: 0.79P=0.37 95% C.I. for  : [-1.21%, 3.21%] 2. Superiority trial

7. Conclusions: Ex I: significant result  SK & rt-PA have different effect Ex II: non-significant result  SK & rt-PA have same effect??? Can we conclude for example II that SK & rt-PA are NOT different? Non-significant result DOES NOT imply that two treatments are equally good(bad) One can NEVER prove that two treatments are equally good (bad) 2. Superiority trial

8. Classical result: P < 0.05  95% C.I. does NOT include  = 0 P > 0.05  95% C.I. includes  = 0 Two-sided 95% C.I.  1-sided 97.5% C.I. Classical test = superiority test Classical trial = superiority trial Assumed  (in sample size calculations) for superiority trial =  S

9. 2. Superiority (trial) Superiority tests

10. 2. Superiority trial E better C better difference + - 30-day mortality rate E - C 0 = 2-sided 95% C.I. H 0 :  = 0 H a :  0 at  Aim superiority trial: Show that E is better than C How? Show that 95% C.I. does not contain 0 = significant at 0.05 In fact only interested in: H 0 :   0 H a :  0 at  /2

11. Take home message 1 A non-significant result NEVER implies that the 2 treatments are EQUALLY GOOD One can NEVER prove that 2 treatments are EQUALLY GOOD If we believe that 2 treatments are EQUALLY GOOD, then another design is needed  EQUIVALENCE TRIAL

12. 3. Equivalence trial Aim: Aim: Prove that E is equally good as C BUT, this can NEVER be done in practice => Practical definition of “equally good” is needed Possible practical definition: 2 treatments do not differ in effect more than a clinically justified value  E => Define interval of clinical equivalence

13. 3. Equivalence trial Interval of clinical equivalence E betterC better difference + - 30-day mortality rate E - C 0 1%-1% EE

14. H 0 :  > 1% or  < -1% H a : -1% <  < 1% E better C better difference + - 30-day mortality rate E - C 0 1%-1% = 2-sided 95% C.I. Aim equivalence trial: Show that E & C are clinically equivalent How? Show that 95% C.I. is INSIDE in interval of therapeutic equivalence 3. Equivalence trial

15. Clinical equivalence is often not the aim of a RCT bioequivalence trials genericdrugoriginal drugMost equivalence trials = bioequivalence trials to compare a generic drug with an original drug to show that they have the “same” PK profile  PK variables C max, C min and AUC must be “close” In bioequivalence trials, often 90% CI is used  E = value such that: “patient will not detect any change in effect when replacing one drug by the other” Noninferiority trials (next) are often (wrongly) called equivalence trials

16. Take home message 2 If you wish to prove that two treatments are  EQUALLY GOOD perform an equivalence trial

17. 4. Non-inferiority trial Introduction 4. Non-inferiority trial Introduction Equivalence trials are not appropriate for therapeutic trials, e.g. if E is clearly superior to C then “equivalence” does not hold.  Prove that E is NOT worse than C ? ? E better than C (superiority, but not believed) ? E equal to C (not possible to prove)  Prove that E is NOT MUCH worse than C!  Define a margin (upper bound) of what can be tolerated  Define an interval of clinical non-inferiority

18. Showing non-inferiority can be of interest because of: Not ethically possible to do a placebo-controlled trial E is not expected to be better than C on primary efficacy endpoint, but is better on secondary endpoints E is not expected to be better than C on primary efficacy endpoint, but is safer E is not expected to be better than C on primary efficacy endpoint, but is cheaper to produce or easier to administer E is not expected to be better than C on primary efficacy endpoint in clinical trial, but compliance will be better outside the clinical trial and hence efficacy better outside the trial 4. Non-inferiority trial Introduction

19. 4. Non-inferiority trial Interval of clinical non-inferiority Interval of clinical non-inferiority E betterC better difference + - 30-day mortality rate E - C 0 1%  NI

20. H 0 :   > 1% H a :  < 1% Aim non-inferiority trial: Show that E is not (much) inferior to C 4. Non-inferiority trial Interval of clinical non-inferiority How? Show that 95% C.I. is inside interval of therapeutic equivalence E better C better difference + - 30-day mortality rate E - C 0 1% = 1-sided 97.5% C.I.

21. 4. Non-inferiority trial Bingham et al. Superiority tests Non-inferiority comparison

22. Superiority P-values Non-inferiority comparison 4. Non-inferiority trial Bingham et al.

23. Two ways to choose the margin  NI : Direct comparison (clinical reasoning): E  C  NI is determined on clinical reasoning Indirect comparison (putative placebo): E  P(lacebo) via C  NI is determined on statistical reasoning Combination: E  C & E  P via C  NI is determined on clinical & statistical reasoning 4. Non-inferiority trial Determination of margin (  NI )

24. Thrombolytic example: Choice of   = 1% can be driven by different reasonings (clinical, statistical, clinical & statistical) Clinical: 1% = largest difference without causing concern Statistical: 1% = difference => safely conclude E better than P Combined: 1% = difference => safely conclude E better than P & without causing concern 4. Non-inferiority trial Example SK versus rt-PA 4. Non-inferiority trial Example SK versus rt-PA Rarely used Often used

25. Determine  NI clinically: Consensus on  NI ? Not easy when YOU are performing the first non-inferiority study in that therapeutic domain (e.g. malaria study) Possible to find a clinically acceptable  NI ? Difficult to justify (purely on clinical grounds) in a mortality trial (e.g. ASSENT II study) 4. Non-inferiority trial Margin determined clinically 4. Non-inferiority trial Margin determined clinically Establishing margin on purely clinical grounds is difficult

26. 4. Non-inferiority trial Bingham et al. 4. Non-inferiority trial Bingham et al. ?

27. Determe  NI Determine difference of C  P by e.g. a meta-analysis => 2% better Determine 95% C.I. around 2% equal to, say, [1.7%, 2.3%] Then E can be at most 1.7% worse than C to guarantee (with 95% confidence) that E is better than P Thus   < 1.7% 4. Non-inferiority trial Margin determined clinically & statistically 4. Non-inferiority trial Margin determined clinically & statistically EC 2% P 95% CI 1.7% 2.3% mortality Check if 1.7% is clinically acceptable, if not lower  NI (to say 1%)

28. Some considerations C must have a well-established, predictable, quantifible effect Multiple placebo-controlled RCTs must be available If not, then there is always the risk that C cannot be “proven” better as P Constancy assumption C  P effect remains the same 4. Non-inferiority trial Choice of Active Control

29. ASSENT II (one of the 1st NI trials in the area) RCT comparing single-bolus tenecteplase (E) with accelerated infusion of alteplase (C) in acute m.i. Primary endpoint = 30-day mortality When  NI = 1%, region of non-inferiority 4. Non-inferiority trial ASSENT II study-1 4. Non-inferiority trial ASSENT II study-1

30. Problem with non-inferiority region: for small mortality rates under alteplase, the allowable relative risk is too high. Let NI margin depend on true alteplase result: change-point at 7.2% (GUSTO III study) 4. Non-inferiority trial ASSENT II study-2 4. Non-inferiority trial ASSENT II study-2

31. How was absolute difference = 1% chosen and how was rr = 1.14 determined? Questions: Determination of margin in collaboration with FDA  NI = 1%: because of GUSTO-1 trial: alteplase was 1% better than SK and SK has proved to be better than placebo + taking 90% confidence intervals into account rr NI = 1.14, a result of the Fibrinolitics Therapy Trialists meta-analysis showing the effect of SK + effect of rt-PA versus SK from GUSTO-1 study + taking 90% confidence intervals into account. 4. Non-inferiority trial ASSENT II study-3 4. Non-inferiority trial ASSENT II study-3

32. Results: 90% C.I. was used instead of 95% C.I. (early NI trial) Endpoint = 30-day mortality E (tenecteplase): 6.16% C (alteplase): 6.18% rr = 0.997 90% C.I. = [0.904, 1.101] Conclusion: E not-inferior to C 4. Non-inferiority trial ASSENT II study-4 4. Non-inferiority trial ASSENT II study-4

33. 4. Non-inferiority trial Malaria study-1 4. Non-inferiority trial Malaria study-1 An open randomized multi-centre clinical trial in Africa, comparing 3 artemisinin-based combination treatments: (1)ASMP (fixed dose over 3 days) (2)ASMP (fixed dose over 24 hours) (3)Artemether-Lumefantrine (AL) (fixed dose over 3 days) on Plasmodium falciparum malaria Main objectives 1.To test the hypothesis that ASMP as fixed dose administered over 24 hours is not inferior in efficacy to the same drug administered over 3 days, measured by the primary endpoint: PCR corrected ACPR on day 28. 2.To test the hypothesis that ASMP as fixed dose is not inferior in efficacy to AL as follows …………. ASMP fixed dose over 24 hours is easier to administer

34. 4. Non-inferiority trial Malaria study-2 4. Non-inferiority trial Malaria study-2 For the first non-inferiority analysis: H 0 : True proportion of cured patients treated with ASMP on 3 days - True proportion of cured patients treated with ASMP on 24 hours  6 % The corresponding alternative hypothesis is: H a : True proportion of cured patients treated with ASMP on 3 days - True proportion of cured patients treated with ASMP on 24 hours < 6 % In early studies with the combination AL, recrudescence of malaria on day 28 was found to be low and varies between 0 and 5%. Re-infection is however sometimes rather high and can vary from 1 – 20 %, particularly in areas with high malaria transmission pressure (Mutabingwa et al., 2005). In some more recent studies, recrudescence was found to be 6 and 8 % respectively (Falade, 2005 and Martensson, 2005). We conclude that, taking into account the studies obtaining a recrudescence of 0 to 5% and the studies mentioning a recrudescence of 6-8%, a non-inferiority interval bounded by 6% can be motivated. Although, the exact choice of the clinical difference is difficult to make.

35. Choice of outcome: difference or ratio? To establish E  P effect, we assumed constancy of C  P effect Constancy assumption involves historical data Absolute risk reduction (  ) is less stable than rr  some (empirical) preference for rr 4. Non-inferiority trial Difference versus ratio 4. Non-inferiority trial Difference versus ratio

36. Analysis population Superiority trial: Standard analysis is based on ITT (intention-to-treat) population Reason = because of conservative effect of ITT approach Non-inferiority trial: ITT analysis is NOT conservative: dropouts and bad conduct of the study push the results of the 2 arms towards each other PP (per-protocol) analysis is preferred but does not provide the ultimate answer Pragmatic approach: do PP & ITT analysis 4. Non-inferiority trial ITT or PP analysis? 4. Non-inferiority trial ITT or PP analysis?

37. Sample size calculations Superiority trial: n depends on (among other things) on  S = the clinically important difference NI trial: n depends on (among other things) on  NI = the upper-bound for non-inferiority When  S =  NI the sample sizes are equal  S for a superiority trial must be greater than  NI in a NI trial  sample size of NI trial > >  sample size of superiority trial 4. Non-inferiority trial Sample size calculations 4. Non-inferiority trial Sample size calculations

38. Non-inferiority & superiority in the same trial Applied to the same population (ITT or PP) Non-inferiority & superiority are tested both at 0.05 (no penalty) because of Closed Testing Principle When non-inferiority is applied to PP & superiority to ITT First non-inferiority & then superiority: no penalty First superiority & then non-inferiority: penalty (multiplicity adjustment) 4. Non-inferiority trial Sample size calculations 4. Non-inferiority trial Sample size calculations

39. A = B & B = C  A = C A < B & B < C  A < C A  B & B  C  A  C ??? A ni B & B ni C  A ni C ??? A ni B & B ni C  A ni P ??? (biocreep) A bit of mathematics

40. Non-significant result with a superiority trial is NOT a proof of equality Goals for the three designs are different: Superiority trial: (say) E is better than C Equivalence trial: E is not too different from C Non-inferiority trial: E is not much worse than C (Equivalence and) non-inferiority depend on choices of the trialist: Interval of clinical (equivalence) non-inferiority 90%  95% C.I. NI trials Make life complicated  if possible use placebo-controlled RCT Unethical ? (Garattini & Bertele, The Lancet, 2007) Take home messages

41. Tips for reading (NI trials) Look carefully at the definition of non-inferiority. This is of crucial importance for the appreciation of the result. Check if definition of non-inferiority is well justified for a clinical viewpoint. When comparing non-inferiority studies, check that definition of NI is the same Check the conduct of the trial. All aspects which reduce the quality of the trial will help “showing” not-inferiority! Non-inferiority CAN NOT be defined/claimed a posteriori!

42. Thank you for your attention